Inversion of lake transparency using remote sensing and deep hybrid recurrent models

Abstract

Utilizing computer technology and remote sensing data, the extraction of water-related features of lakes has become a hot topic in lake ecological research. Addressing challenges like the high optical complexity of lake water bodies, the inadequacy of samples for capturing complex optical properties, and the difficulty of large - scale application of simplified lake water optical models, this study robustly integrated LSTM and GRU network structures to construct an accurate and efficient lake water transparency inversion model (WTIM). The model utilized Landsat - 8 remote sensing data, field measurements, and simulated data to form a sample set. This model is specifically designed for rapid, large-scale, and automated remote sensing inversion of lake transparency. The results show that the WTIM model can invert lake water transparency with good accuracy (R²=0.78, MAE=0.64, RMSE=0.84, MAPE=52.31 %), and the model has excellent robustness. Analysis of the time series characteristics of Chinese lakes from 2014 to 2021 reveals that lake water transparency in China first decreased and then increased over time, showing an overall decreasing trend. Analysis of spatial variation characteristics indicates that lake transparency in the Qinghai-Tibet Plateau lake region is increasing, mainly due to the inflow of glacial meltwater into lakes caused by global warming. In contrast, lake transparency in the eastern plain lake region and the northeast plain lake region is decreasing, likely due to intense human industrial and agricultural activities. Our research can provide a reference for lake transparency inversion.